diff --git a/src/conf.h b/src/conf.h
index cb83c5b4..cb9f8581 100644
--- a/src/conf.h
+++ b/src/conf.h
@@ -460,6 +460,7 @@ private:
 #define UPX_F_MACH_i386         29
 #define UPX_F_LINUX_ELF32_MIPSEL 30
 #define UPX_F_VMLINUZ_ARMEL     31
+#define UPX_F_MACH_ARMEL        32
 
 #define UPX_F_PLAIN_TEXT        127
 
diff --git a/src/p_mach.cpp b/src/p_mach.cpp
index 6dd774a3..8f073ca1 100644
--- a/src/p_mach.cpp
+++ b/src/p_mach.cpp
@@ -44,6 +44,11 @@ static const
 static const
 #include "stub/i386-darwin.macho-fold.h"
 
+static const
+#include "stub/arm-darwin.macho-entry.h"
+static const
+#include "stub/arm-darwin.macho-fold.h"
+
 template <class T>
 PackMachBase<T>::PackMachBase(InputFile *f, unsigned cputype, unsigned flavor,
         unsigned count, unsigned size) :
@@ -69,6 +74,11 @@ const int *PackMachBase<T>::getCompressionMethods(int method, int level) const
     return Packer::getDefaultCompressionMethods_le32(method, level);
 }
 
+const int *PackMachARMEL::getCompressionMethods(int method, int level) const
+{
+    return Packer::getDefaultCompressionMethods_8(method, level);
+}
+
 
 const int *PackMachPPC32::getFilters() const
 {
@@ -82,6 +92,12 @@ int const *PackMachI386::getFilters() const
     return filters;
 }
 
+int const *PackMachARMEL::getFilters() const
+{
+    static const int filters[] = { 0x50, FT_END };
+    return filters;
+}
+
 Linker *PackMachPPC32::newLinker() const
 {
     return new ElfLinkerPpc32;
@@ -92,6 +108,11 @@ Linker *PackMachI386::newLinker() const
     return new ElfLinkerX86;
 }
 
+Linker *PackMachARMEL::newLinker() const
+{
+    return new ElfLinkerArmLE;
+}
+
 template <class T>
 void
 PackMachBase<T>::addStubEntrySections(Filter const *)
@@ -171,6 +192,21 @@ void PackMachI386::addStubEntrySections(Filter const *ft)
     addLoader("FOLDEXEC", NULL);
 }
 
+void PackMachARMEL::addStubEntrySections(Filter const */*ft*/)
+{
+    addLoader("MACHMAINX", NULL);
+   //addLoader(getDecompressorSections(), NULL);
+    addLoader(
+        ( M_IS_NRV2E(ph.method) ? "NRV_HEAD,NRV2E,NRV_TAIL"
+        : M_IS_NRV2D(ph.method) ? "NRV_HEAD,NRV2D,NRV_TAIL"
+        : M_IS_NRV2B(ph.method) ? "NRV_HEAD,NRV2B,NRV_TAIL"
+        : M_IS_LZMA(ph.method)  ? "LZMA_ELF00,+80C,LZMA_DEC20,LZMA_DEC30"
+        : NULL), NULL);
+    if (hasLoaderSection("CFLUSH"))
+        addLoader("CFLUSH");
+    addLoader("MACHMAINY,IDENTSTR,+40,MACHMAINZ,FOLDEXEC", NULL);
+}
+
 
 template <class T>
 void PackMachBase<T>::defineSymbols(Filter const *)
@@ -217,7 +253,8 @@ PackMachBase<T>::buildMachLoader(
     delete [] cprLoader;
 
     int const GAP = 128;  // must match stub/l_mac_ppc.S
-    segcmdo.vmsize += h.sz_unc - h.sz_cpr + GAP + 64;
+    int const NO_LAP = 64;  // must match stub/src/*darwin*.S
+    segcmdo.vmsize += h.sz_unc - h.sz_cpr + GAP + NO_LAP;
 
     addStubEntrySections(ft);
 
@@ -241,6 +278,14 @@ PackMachI386::buildLoader(const Filter *ft)
         stub_i386_darwin_macho_fold,  sizeof(stub_i386_darwin_macho_fold),  ft );
 }
 
+void
+PackMachARMEL::buildLoader(const Filter *ft)
+{
+    buildMachLoader(
+        stub_arm_darwin_macho_entry, sizeof(stub_arm_darwin_macho_entry),
+        stub_arm_darwin_macho_fold,  sizeof(stub_arm_darwin_macho_fold),  ft );
+}
+
 template <class T>
 void PackMachBase<T>::patchLoader() { }
 
@@ -305,6 +350,20 @@ void PackMachI386::pack4(OutputFile *fo, Filter &ft)  // append PackHeader
     fo->rewrite(&linfo, sizeof(linfo));
 }
 
+void PackMachARMEL::pack4(OutputFile *fo, Filter &ft)  // append PackHeader
+{
+    // offset of p_info in compressed file
+    overlay_offset = sizeof(mhdro) + sizeof(segcmdo) + sizeof(threado) + sizeof(linfo);
+
+    super::pack4(fo, ft);
+    segcmdo.filesize = fo->getBytesWritten();
+    segcmdo.vmsize += segcmdo.filesize;
+    fo->seek(sizeof(mhdro), SEEK_SET);
+    fo->rewrite(&segcmdo, sizeof(segcmdo));
+    fo->rewrite(&threado, sizeof(threado));
+    fo->rewrite(&linfo, sizeof(linfo));
+}
+
 void PackMachPPC32::pack3(OutputFile *fo, Filter &ft)  // append loader
 {
     BE32 disp;
@@ -333,6 +392,20 @@ void PackMachI386::pack3(OutputFile *fo, Filter &ft)  // append loader
     super::pack3(fo, ft);
 }
 
+void PackMachARMEL::pack3(OutputFile *fo, Filter &ft)  // append loader
+{
+    LE32 disp;
+    unsigned const zero = 0;
+    unsigned len = fo->getBytesWritten();
+    fo->write(&zero, 3& (0u-len));
+    len += (3& (0u-len));
+    disp = len - sz_mach_headers;
+    fo->write(&disp, sizeof(disp));
+
+    threado.state.pc = len + sizeof(disp) + segcmdo.vmaddr;  /* entry address */
+    super::pack3(fo, ft);
+}
+
 // Determine length of gap between PT_LOAD phdri[k] and closest PT_LOAD
 // which follows in the file (or end-of-file).  Optimize for common case
 // where the PT_LOAD are adjacent ascending by .p_offset.  Assume no overlap.
@@ -480,6 +553,16 @@ void PackMachI386::pack1_setup_threado(OutputFile *const fo)
     fo->write(&threado, sizeof(threado));
 }
 
+void PackMachARMEL::pack1_setup_threado(OutputFile *const fo)
+{
+    threado.cmd = Mach_segment_command::LC_UNIXTHREAD;
+    threado.cmdsize = sizeof(threado);
+    threado.flavor = my_thread_flavor;
+    threado.count =  my_thread_state_word_count;
+    memset(&threado.state, 0, sizeof(threado.state));
+    fo->write(&threado, sizeof(threado));
+}
+
 template <class T>
 void PackMachBase<T>::pack1(OutputFile *const fo, Filter &/*ft*/)  // generate executable header
 {
diff --git a/src/p_mach.h b/src/p_mach.h
index d75f5224..842702f0 100644
--- a/src/p_mach.h
+++ b/src/p_mach.h
@@ -197,6 +197,19 @@ struct Mach_i386_new_thread_state
 }
 __attribute_packed;
 
+template <class TMachITypes>
+struct Mach_ARM_thread_state
+{
+    typedef typename TMachITypes::Word Word;
+
+    Word r[13];  // r0-r12
+    Word sp;  // r13
+    Word lr;  // r14
+    Word pc;  // r15
+    Word cpsr;
+}
+__attribute_packed;
+
 }  // namespace N_Mach
 
 namespace N_Mach32 {
@@ -253,6 +266,7 @@ struct MachClass_32
     typedef N_Mach::Mach_section_command<MachITypes> Mach_section_command;
     typedef N_Mach::Mach_ppc_thread_state<MachITypes> Mach_ppc_thread_state;
     typedef N_Mach::Mach_i386_thread_state<MachITypes> Mach_i386_thread_state;
+    typedef N_Mach::Mach_ARM_thread_state<MachITypes> Mach_ARM_thread_state;
 
     static void compileTimeAssertions() {
         BeLePolicy::compileTimeAssertions();
@@ -311,6 +325,7 @@ typedef MachClass_LE64::Mach_section_command   MachLE64_section_command;
 
 typedef MachClass_BE32::Mach_ppc_thread_state  Mach_ppc_thread_state;
 typedef MachClass_LE32::Mach_i386_thread_state Mach_i386_thread_state;
+typedef MachClass_LE32::Mach_ARM_thread_state  Mach_ARM_thread_state;
 #include "p_unix.h"
 
 
@@ -491,6 +506,43 @@ protected:
     Mach_thread_command threado;
 };
 
+class PackMachARMEL : public PackMachBase<MachClass_LE32>
+{
+    typedef PackMachBase<MachClass_LE32> super;
+
+public:
+    PackMachARMEL(InputFile *f) : super(f, Mach_header::CPU_TYPE_ARM,
+        (unsigned)Mach_thread_command::ARM_THREAD_STATE,
+        sizeof(Mach_ARM_thread_state)>>2, sizeof(threado)) { }
+
+    virtual int getFormat() const { return UPX_F_MACH_ARMEL; }
+    virtual const char *getName() const { return "Mach/ARMEL"; }
+    virtual const char *getFullName(const options_t *) const { return "ARMEL-darwin.macho"; }
+protected:
+    virtual const int *getCompressionMethods(int method, int level) const;
+    virtual const int *getFilters() const;
+
+    virtual void pack1_setup_threado(OutputFile *const fo);
+    virtual void pack3(OutputFile *, Filter &);  // append loader
+    virtual void pack4(OutputFile *, Filter &);  // append PackHeader
+    virtual Linker* newLinker() const;
+    virtual void buildLoader(const Filter *ft);
+    virtual void addStubEntrySections(Filter const *);
+
+    struct Mach_thread_command
+    {
+        LE32 cmd;            /* LC_THREAD or  LC_UNIXTHREAD */
+        LE32 cmdsize;        /* total size of this command */
+        LE32 flavor;
+        LE32 count;          /* sizeof(following_thread_state)/4 */
+        Mach_ARM_thread_state state;
+    #define WANT_MACH_THREAD_ENUM
+    #include "p_mach_enum.h"
+    }
+    __attribute_packed;
+    Mach_thread_command threado;
+};
+
 class PackMachFat : public Packer
 {
     typedef Packer super;
diff --git a/src/p_mach_enum.h b/src/p_mach_enum.h
index cb2120b6..54ee8fc9 100644
--- a/src/p_mach_enum.h
+++ b/src/p_mach_enum.h
@@ -40,9 +40,15 @@
     };
     enum { // cputype
         CPU_TYPE_I386      =          7,
+        CPU_TYPE_ARM       =         12,
         CPU_TYPE_POWERPC   = 0x00000012,
         CPU_TYPE_POWERPC64 = 0x01000012
     };
+    enum { // cpusubtype
+        CPU_SUBTYPE_ARM_ALL = 0,
+        CPU_SUBTYPE_ARM_V4T = 5,
+        CPU_SUBTYPE_ARM_V6  = 6
+    };
     enum { // filetype
         MH_EXECUTE = 2
     };
@@ -98,7 +104,8 @@
     enum { // thread flavor
         PPC_THREAD_STATE = 1,
         i386_THREAD_STATE = 1,
-        i386_OLD_THREAD_STATE = -1
+        i386_OLD_THREAD_STATE = -1,
+        ARM_THREAD_STATE = 1
     };
 #endif  /*}*/
 
diff --git a/src/packmast.cpp b/src/packmast.cpp
index f78ef1e0..8e717d9d 100644
--- a/src/packmast.cpp
+++ b/src/packmast.cpp
@@ -278,6 +278,8 @@ Packer* PackMaster::visitAllPackers(visit_func_t func, InputFile *f, const optio
         return p;
     if ((p = func(new PackMachI386(f), user)) != NULL)
         return p;
+    if ((p = func(new PackMachARMEL(f), user)) != NULL)
+        return p;
 
     return NULL;
 }
diff --git a/src/stub/Makefile b/src/stub/Makefile
index c0a3a522..5785e929 100644
--- a/src/stub/Makefile
+++ b/src/stub/Makefile
@@ -52,6 +52,8 @@ STUBS += amd64-linux.kernel.vmlinux.h
 STUBS += amd64-linux.kernel.vmlinux-head.h
 STUBS += armel-eabi-linux.elf-entry.h
 STUBS += armel-eabi-linux.elf-fold.h
+STUBS += arm-darwin.macho-entry.h
+STUBS += arm-darwin.macho-fold.h
 STUBS += arm-linux.elf-entry.h
 STUBS += arm-linux.elf-fold.h
 STUBS += arm-linux.kernel.vmlinux.h
@@ -327,6 +329,35 @@ tmp/armel-linux.elf-main.o : $(srcdir)/src/$$T.c
 	$(call tc,f-objstrip,$@)
 	$(call tc,objdump) -dr $(tc_objdump_disasm_options) $@ | $(RTRIM) > $@.disasm
 
+# /***********************************************************************
+# // arm-darwin.macho
+# ************************************************************************/
+
+# info: we use the tc settings from arm-linux.elf, but override v4 with v5
+arm-darwin.macho%.h : tc_list = arm-linux.elf default
+arm-darwin.macho%.h : tc_bfdname = elf32-littlearm
+tc.arm-darwin.macho-entry.gcc  = arm-linux-gcc-4.1.0 -march=armv5 -nostdinc -MMD -MT $@
+tc.arm-darwin.macho-fold.gcc   = arm-linux-gcc-4.1.0 -march=armv5 -nostdinc -MMD -MT $@
+
+arm-darwin.macho-entry.h : $(srcdir)/src/$$T.S
+	$(call tc,gcc) -c $< -o tmp/$T.bin
+	$(call tc,f-embed_objinfo,tmp/$T.bin)
+	$(call tc,bin2h) tmp/$T.bin $@
+
+arm-darwin.macho-fold.h : tmp/$$T.o tmp/arm-darwin.macho-main.o
+	$(call tc,ld) --no-warn-mismatch --strip-all --oformat binary -Map tmp/$T.map $(filter %.o,$^) -o tmp/$T.bin
+	chmod a-x tmp/$T.bin
+	$(call tc,bin2h) tmp/$T.bin $@
+
+tmp/arm-darwin.macho-fold.o : $(srcdir)/src/$$T.S
+	$(call tc,gcc) -c $< -o $@
+	$(call tc,f-objstrip,$@)
+
+tmp/arm-darwin.macho-main.o : $(srcdir)/src/$$T.c
+	$(call tc,gcc) -c -Os $< -o $@
+	$(call tc,f-objstrip,$@)
+	$(call tc,objdump) -dr $(tc_objdump_disasm_options) $@ | $(RTRIM) > $@.disasm
+
 # /***********************************************************************
 # // arm-linux.elf
 # ************************************************************************/
diff --git a/src/stub/i386-darwin.macho-fold.h b/src/stub/i386-darwin.macho-fold.h
index 8b301589..0783ba62 100644
--- a/src/stub/i386-darwin.macho-fold.h
+++ b/src/stub/i386-darwin.macho-fold.h
@@ -1,5 +1,5 @@
 /* i386-darwin.macho-fold.h
-   created from i386-darwin.macho-fold.bin, 1105 (0x451) bytes
+   created from i386-darwin.macho-fold.bin, 1096 (0x448) bytes
 
    This file is part of the UPX executable compressor.
 
@@ -31,13 +31,13 @@
  */
 
 
-#define STUB_I386_DARWIN_MACHO_FOLD_SIZE    1105
-#define STUB_I386_DARWIN_MACHO_FOLD_ADLER32 0x5ad106a1
-#define STUB_I386_DARWIN_MACHO_FOLD_CRC32   0x01a4878b
+#define STUB_I386_DARWIN_MACHO_FOLD_SIZE    1096
+#define STUB_I386_DARWIN_MACHO_FOLD_ADLER32 0xee0f0416
+#define STUB_I386_DARWIN_MACHO_FOLD_CRC32   0xa5096bf9
 
-unsigned char stub_i386_darwin_macho_fold[1105] = {
+unsigned char stub_i386_darwin_macho_fold[1096] = {
 /* 0x0000 */ 106,  0,137,231,141,117,  2,139, 19,137,217, 41,209,139, 89, 24,
-/* 0x0010 */ 184,  0,  8,  0,  0, 57,216,118,  2,137,195, 41,220, 96,232,  6,
+/* 0x0010 */ 184,  0,  8,  0,  0, 57,216,118,  2,137,195, 41,220, 96,232,  1,
 /* 0x0020 */   3,  0,  0,139, 76, 36, 16,141,100, 12, 32,255, 96, 40,139, 68,
 /* 0x0030 */  36,  4,139, 76, 36,  8,139, 16, 15,202,137, 16,131,233,  4,141,
 /* 0x0040 */  64,  4,115,242,195, 90, 15, 52,176,  1,235,  2,176, 74,235,  2,
@@ -64,46 +64,45 @@ unsigned char stub_i386_darwin_macho_fold[1105] = {
 /* 0x0190 */ 228,139, 85, 12,199, 69,208,  0,  0,  0,  0,137, 69,224,139, 69,
 /* 0x01a0 */  20,137, 85,220,139, 85, 24,137, 69,216,139, 93,232,139, 69,232,
 /* 0x01b0 */ 137, 85,212, 49,210,131,195, 28,199, 69,204,  0,  0,  0,  0, 59,
-/* 0x01c0 */  80, 16, 15,131, 86,  1,  0,  0,139,  3,131,248,  1, 15,133, 29,
+/* 0x01c0 */  80, 16, 15,131, 81,  1,  0,  0,139,  3,131,248,  1, 15,133, 24,
 /* 0x01d0 */   1,  0,  0,139, 83, 24,139, 67, 28,139, 75, 36,137,214,  1,208,
 /* 0x01e0 */ 137, 85,240,137, 69,200,137,208, 37,255, 15,  0,  0,137,207, 41,
-/* 0x01f0 */ 198,  1,199,137, 77,236,116, 69,139, 69,228, 49,210,  3, 67, 32,
-/* 0x0200 */  82,133,201, 80,139, 69,220,117,  3,131,200,255,131,125,224,  0,
-/* 0x0210 */  80,117,  9,133,201,184, 18,  0,  0,  0,117,  5,184, 18, 16,  0,
-/* 0x0220 */   0,131,125,224,  0, 80,106,  3,137,248,116,  3,141, 71,  3, 80,
-/* 0x0230 */  86,232, 42,254,255,255,131,196, 28, 57,198,117, 94,131,125,224,
-/* 0x0240 */   0,116, 42,131,123, 36,  0,116, 36,131,123, 32,  0,117, 11,131,
-/* 0x0250 */ 125, 16,  0,116,  5,139, 85, 16,137, 50,255,117,212,255,117,216,
-/* 0x0260 */ 141, 85,236,139, 69,224,232, 65,254,255,255, 88, 90,137,248,141,
-/* 0x0270 */  20, 62,247,216, 37,255, 15,  0,  0,137, 69,196,116,  8,137,193,
-/* 0x0280 */ 198,  2,  0, 66,226,250,133,255,116, 24,255,115, 44, 87, 86,232,
-/* 0x0290 */ 184,253,255,255,131,196, 12,133,192,116,  7,106,127,232,166,253,
-/* 0x02a0 */ 255,255,139, 85,196,141,  4, 23,  1,198, 59,117,200,115, 35,106,
-/* 0x02b0 */   0,106,  0,106,255,104, 18, 16,  0,  0,255,115, 44, 41,117,200,
-/* 0x02c0 */ 255,117,200, 86,232,151,253,255,255,131,196, 28, 57,198,116, 58,
-/* 0x02d0 */ 235,201,131,125,224,  0,116, 50,141, 71,  3, 37,255, 15,  0,  0,
-/* 0x02e0 */ 131,248,  3,119, 37, 80, 86,232,100,253,255,255, 89, 94,235, 26,
-/* 0x02f0 */ 131,232,  4,131,248,  1,119, 18,131,123,  8,  1,117, 12,131,123,
-/* 0x0300 */  12, 16,117,  6,141, 67, 16,137, 69,208,255, 69,204,139, 85,232,
-/* 0x0310 */ 139, 69,204,  3, 91,  4, 59, 66, 16,233,164,254,255,255,139, 69,
-/* 0x0320 */ 208,141,101,244, 91, 94, 95,201,195, 85,137,229, 87, 86, 83,131,
-/* 0x0330 */ 236, 32,199, 69,212,  0,  0,  0,  0,139, 85, 32,139, 69, 24,139,
-/* 0x0340 */  93, 16,137, 69,216,139,117, 20,141, 66, 24,137,117,232,137, 69,
-/* 0x0350 */ 240,139, 69, 28,131,232, 24,137, 69,236,139, 66, 24,139, 85,240,
-/* 0x0360 */ 106,  0,137, 69,228,139, 69,236,137, 85,224,137, 69,220,141, 85,
-/* 0x0370 */ 228,141, 69,236, 83,232, 50,253,255,255,255,117, 12, 83, 49,210,
-/* 0x0380 */ 255,117,  8,141, 69,220,106,255, 80,137,240,232,239,253,255,255,
-/* 0x0390 */  49,210,137,195,141, 70, 28,131,196, 28,139, 78, 16, 57,202, 15,
-/* 0x03a0 */ 131,162,  0,  0,  0,131, 56, 14, 15,133,144,  0,  0,  0,  3, 64,
-/* 0x03b0 */   8,106,  0,106,  0, 80,232,161,252,255,255,131,196, 12,133,192,
-/* 0x03c0 */ 137,199,120, 25, 49,210,139, 69,212, 82, 80,255,117,216, 86, 87,
-/* 0x03d0 */ 232,127,252,255,255,131,196, 20, 57, 69,216,116, 15,106,127,232,
-/* 0x03e0 */ 100,252,255,255,139, 91,  8,137, 93,212,235,216,129, 62,202,254,
-/* 0x03f0 */ 186,190,117, 42, 15,182, 70,  7,141, 94,  8,107,192, 20,131,192,
-/* 0x0400 */   8, 80, 86,232, 38,252,255,255, 89, 90, 49,192,139, 86,  4, 57,
-/* 0x0410 */ 208,115, 11,131, 59,  7,116,204, 64,131,195, 20,235,241,106,  0,
-/* 0x0420 */ 139, 85,212,106,  0,137,240,106,  0, 87,106,  0,232, 78,253,255,
-/* 0x0430 */ 255, 87,137,195,232, 31,252,255,255,131,196, 24,235,  9,  3, 64,
-/* 0x0440 */   4, 66,233, 86,255,255,255,141,101,244,137,216, 91, 94, 95,201,
-/* 0x0450 */ 195
+/* 0x01f0 */ 198,  1,199,137, 77,236,116, 66,139, 69,228,  3, 67, 32,133,201,
+/* 0x0200 */  80,139, 69,220,117,  3,131,200,255,131,125,224,  0, 80,117,  9,
+/* 0x0210 */ 133,201,184, 18,  0,  0,  0,117,  5,184, 18, 16,  0,  0,131,125,
+/* 0x0220 */ 224,  0, 80,106,  3,137,248,116,  3,141, 71,  3, 80, 86,232, 45,
+/* 0x0230 */ 254,255,255,131,196, 24, 57,198,117, 94,131,125,224,  0,116, 42,
+/* 0x0240 */ 131,123, 36,  0,116, 36,131,123, 32,  0,117, 11,131,125, 16,  0,
+/* 0x0250 */ 116,  5,139, 85, 16,137, 50,255,117,212,255,117,216,141, 85,236,
+/* 0x0260 */ 139, 69,224,232, 68,254,255,255, 88, 90,137,248,141, 20, 62,247,
+/* 0x0270 */ 216, 37,255, 15,  0,  0,137, 69,196,116,  8,137,193,198,  2,  0,
+/* 0x0280 */  66,226,250,133,255,116, 24,255,115, 44, 87, 86,232,187,253,255,
+/* 0x0290 */ 255,131,196, 12,133,192,116,  7,106,127,232,169,253,255,255,139,
+/* 0x02a0 */  85,196,141,  4, 23,  1,198, 59,117,200,115, 33,106,  0,106,255,
+/* 0x02b0 */ 104, 18, 16,  0,  0,255,115, 44, 41,117,200,255,117,200, 86,232,
+/* 0x02c0 */ 156,253,255,255,131,196, 24, 57,198,116, 58,235,203,131,125,224,
+/* 0x02d0 */   0,116, 50,141, 71,  3, 37,255, 15,  0,  0,131,248,  3,119, 37,
+/* 0x02e0 */  80, 86,232,105,253,255,255, 89, 94,235, 26,131,232,  4,131,248,
+/* 0x02f0 */   1,119, 18,131,123,  8,  1,117, 12,131,123, 12, 16,117,  6,141,
+/* 0x0300 */  67, 16,137, 69,208,255, 69,204,139, 85,232,139, 69,204,  3, 91,
+/* 0x0310 */   4, 59, 66, 16,233,169,254,255,255,139, 69,208,141,101,244, 91,
+/* 0x0320 */  94, 95,201,195, 85,137,229, 87, 86, 83,131,236, 32,199, 69,212,
+/* 0x0330 */   0,  0,  0,  0,139, 85, 32,139, 69, 24,139, 93, 16,137, 69,216,
+/* 0x0340 */ 139,117, 20,141, 66, 24,137,117,232,137, 69,240,139, 69, 28,131,
+/* 0x0350 */ 232, 24,137, 69,236,139, 66, 24,139, 85,240,106,  0,137, 69,228,
+/* 0x0360 */ 139, 69,236,137, 85,224,137, 69,220,141, 85,228,141, 69,236, 83,
+/* 0x0370 */ 232, 55,253,255,255,255,117, 12, 83, 49,210,255,117,  8,141, 69,
+/* 0x0380 */ 220,106,255, 80,137,240,232,244,253,255,255, 49,210,137,195,141,
+/* 0x0390 */  70, 28,131,196, 28,139, 78, 16, 57,202, 15,131,158,  0,  0,  0,
+/* 0x03a0 */ 131, 56, 14, 15,133,140,  0,  0,  0,  3, 64,  8,106,  0,106,  0,
+/* 0x03b0 */  80,232,166,252,255,255,131,196, 12,133,192,137,199,120, 21,255,
+/* 0x03c0 */ 117,212,255,117,216, 86, 87,232,136,252,255,255,131,196, 16, 57,
+/* 0x03d0 */  69,216,116, 15,106,127,232,109,252,255,255,139, 91,  8,137, 93,
+/* 0x03e0 */ 212,235,220,129, 62,202,254,186,190,117, 42, 15,182, 70,  7,141,
+/* 0x03f0 */  94,  8,107,192, 20,131,192,  8, 80, 86,232, 47,252,255,255, 89,
+/* 0x0400 */  90, 49,192,139, 86,  4, 57,208,115, 11,131, 59,  7,116,204, 64,
+/* 0x0410 */ 131,195, 20,235,241,106,  0,139, 85,212,106,  0,137,240,106,  0,
+/* 0x0420 */  87,106,  0,232, 87,253,255,255, 87,137,195,232, 40,252,255,255,
+/* 0x0430 */ 131,196, 24,235,  9,  3, 64,  4, 66,233, 90,255,255,255,141,101,
+/* 0x0440 */ 244,137,216, 91, 94, 95,201,195
 };
diff --git a/src/stub/src/arch/arm/v4a/lzma_d.S b/src/stub/src/arch/arm/v4a/lzma_d.S
index 8f970754..335421a5 100644
--- a/src/stub/src/arch/arm/v4a/lzma_d.S
+++ b/src/stub/src/arch/arm/v4a/lzma_d.S
@@ -45,7 +45,7 @@
 #define M_LZMA          14
         ldrb ip,meth; cmp ip,#M_LZMA; bne not_lzma
 
-#if defined(LINUX_ARM_CACHEFLUSH)  /*{*/
+#if defined(LINUX_ARM_CACHEFLUSH)||defined(DARWIN_ARM_CACHEFLUSH)  /*{*/
 	PUSH {dst,ldst, fp,lr}  // dst,ldst for cache flush
 #else  /*}{*/
 	PUSH {          fp,lr}
@@ -116,6 +116,14 @@
         do_sys2 __ARM_NR_cacheflush  // decompressed region
         mov r0,r3  // result value
 #endif  /*}*/
+#if defined(DARWIN_ARM_CACHEFLUSH)  /*{*/
+        mov r4,r0  // save result value
+        POP {r0,r1}  // dst, ldst
+        ldr r1,[r1]  // ldst by reference
+        PUSH {r0,r1}; do_dcache_flush
+        POP  {r0,r1}; do_icache_invalidate
+        mov r0,r4  // result value
+#endif  /*}*/
 
         POP {fp,pc}
 1:
diff --git a/src/stub/src/arch/arm/v4a/macros.S b/src/stub/src/arch/arm/v4a/macros.S
index 1e5f0594..9f3d6b38 100644
--- a/src/stub/src/arch/arm/v4a/macros.S
+++ b/src/stub/src/arch/arm/v4a/macros.S
@@ -34,7 +34,16 @@
                 .section \name
 .endm
 
-#if defined(ARMEL_EABI4)  /*{*/
+#if defined(ARMEL_DARWIN)  /*{*/
+__NR_SYSCALL_BASE = 0
+.macro          do_sys N
+                mov ip,#\N
+                swi 0x80  // returns Carry iff error
+                orrcs r0,r0,#(1<<31)  // force negative on error; FIXME: needed?
+                ret
+.endm
+
+#elif defined(ARMEL_EABI4)  /*}{*/
 
 __NR_SYSCALL_BASE = 0
 .macro          do_sys N
@@ -51,7 +60,7 @@ __NR_SYSCALL_BASE = 0
                 mov r7,r12  // restore r7 from ip
 .endm
 
-#else  /*}{*/
+#elif defined(ARM_OLDABI)  /*}{*/
 
 __NR_SYSCALL_BASE = 0x900000
 .macro          do_sys N
@@ -61,10 +70,17 @@ __NR_SYSCALL_BASE = 0x900000
                 swi \N
 .endm
 
+#else  /*}{*/
+.macro          do_sys N
+                error ARM_OLDABI, ARMEL_EABI4, ARMEL_DARWIN ?
+.endm
+.macro          do_sys2 N
+                error ARM_OLDABI, ARMEL_EABI4, ARMEL_DARWIN ?
+.endm
 #endif  /*}*/
 
 .macro          ret
-                mov pc,lr  /* armv4 fails for thumb interworking */
+                mov pc,lr  /* armv4 lacks 'bx'; fails for thumb interworking */
 .endm
 
 // vi:ts=8:et:nowrap
diff --git a/src/stub/src/arch/arm/v4a/nrv2b_d8.S b/src/stub/src/arch/arm/v4a/nrv2b_d8.S
index 3448e6d6..e45d2223 100644
--- a/src/stub/src/arch/arm/v4a/nrv2b_d8.S
+++ b/src/stub/src/arch/arm/v4a/nrv2b_d8.S
@@ -78,6 +78,14 @@ eof_n2b:
         do_sys2 __ARM_NR_cacheflush  // decompressed region
         mov r0,r4  // result value
 #endif  /*}*/
+#if defined(DARWIN_ARM_CACHEFLUSH)  /*{*/
+        mov r4,r0  // save result value
+        mov r0,r3  // orig_dst
+        mov r1,dst  // dst_len
+        PUSH {r0,r1}; do_dcache_flush
+        POP  {r0,r1}; do_icache_invalidate
+        mov r0,r4  // result value
+#endif  /*}*/
 
         POP {r4,r5, pc}  // return
 
diff --git a/src/stub/src/arch/arm/v4a/nrv2d_d8.S b/src/stub/src/arch/arm/v4a/nrv2d_d8.S
index 9eb048d2..d75d035c 100644
--- a/src/stub/src/arch/arm/v4a/nrv2d_d8.S
+++ b/src/stub/src/arch/arm/v4a/nrv2d_d8.S
@@ -118,6 +118,14 @@ eof_n2d:
         do_sys2 __ARM_NR_cacheflush  // decompressed region
         mov r0,r4  // result value
 #endif  /*}*/
+#if defined(DARWIN_ARM_CACHEFLUSH)  /*{*/
+        mov r4,r0  // save result value
+        mov r0,r3  // orig_dst
+        mov r1,dst  // dst_len
+        PUSH {r0,r1}; do_dcache_flush
+        POP  {r0,r1}; do_icache_invalidate
+        mov r0,r4  // result value
+#endif  /*}*/
 
         POP {r4,r5,r6,r7 ,pc}
 
diff --git a/src/stub/src/arch/arm/v4a/nrv2e_d8.S b/src/stub/src/arch/arm/v4a/nrv2e_d8.S
index f1c834b3..6aaa824d 100644
--- a/src/stub/src/arch/arm/v4a/nrv2e_d8.S
+++ b/src/stub/src/arch/arm/v4a/nrv2e_d8.S
@@ -118,6 +118,14 @@ eof_n2e:
         do_sys2 __ARM_NR_cacheflush  // decompressed region
         mov r0,r4  // result value
 #endif  /*}*/
+#if defined(DARWIN_ARM_CACHEFLUSH)  /*{*/
+        mov r4,r0  // save result value
+        mov r0,r3  // orig_dst
+        mov r1,dst  // dst_len
+        PUSH {r0,r1}; do_dcache_flush
+        POP  {r0,r1}; do_icache_invalidate
+        mov r0,r4  // result value
+#endif  /*}*/
 
         POP {r4,r5,r6,r7 ,pc}
 
diff --git a/src/stub/src/arch/arm/v5a/macros.S b/src/stub/src/arch/arm/v5a/macros.S
index 3ef210ed..e4228493 100644
--- a/src/stub/src/arch/arm/v5a/macros.S
+++ b/src/stub/src/arch/arm/v5a/macros.S
@@ -34,7 +34,27 @@
                 .section \name
 .endm
 
-#if defined(ARMEL_EABI4)  /*{*/
+#if defined(ARMEL_DARWIN)  /*{*/
+__NR_SYSCALL_BASE = 0
+.macro          do_sys N
+                mov ip,#\N
+                swi 0x80  // sets Carry iff error
+                orrcs r0,r0,#(1<<31)  // force negative on error; FIXME: needed?
+                ret
+.endm
+
+.macro          do_dcache_flush  // In: r0=addr; r1=len
+                mov r3,#1  // _sys_dcache_flush
+                mov ip,#(1<<31)  // syscall number?
+                swi 0x80
+.endm
+
+.macro          do_icache_invalidate  // In: r0=addr; r1=len
+                mov r3,#0  // _sys_icache_invalidate
+                mov ip,#(1<<31)  // syscall number?
+                swi 0x80
+.endm
+#elif defined(ARMEL_EABI4)  /*}{*/
 
 __NR_SYSCALL_BASE = 0
 .macro          do_sys N
@@ -51,7 +71,7 @@ __NR_SYSCALL_BASE = 0
                 mov r7,r12  // restore r7 from ip
 .endm
 
-#else  /*}{*/
+#elif defined(ARM_OLDABI)  /*}{*/
 
 __NR_SYSCALL_BASE = 0x900000
 .macro          do_sys N
@@ -61,6 +81,13 @@ __NR_SYSCALL_BASE = 0x900000
                 swi \N
 .endm
 
+#else  /*}{*/
+.macro          do_sys N
+                error \N  // ARM_OLDABI, ARMEL_EABI4, ARMEL_DARWIN ?
+.endm
+.macro          do_sys2 N
+                error \N  // ARM_OLDABI, ARMEL_EABI4, ARMEL_DARWIN ?
+.endm
 #endif  /*}*/
 
 .macro          ret
diff --git a/src/stub/src/arm-linux.elf-entry.S b/src/stub/src/arm-linux.elf-entry.S
index be39a047..89e48706 100644
--- a/src/stub/src/arm-linux.elf-entry.S
+++ b/src/stub/src/arm-linux.elf-entry.S
@@ -29,6 +29,7 @@
 *  <jreiser@users.sourceforge.net>
 */
 
+#define ARM_OLDABI 1
 #include "arch/arm/v4a/macros.S"
 
 #define bkpt .long 0xe1200070
diff --git a/src/stub/src/arm-linux.elf-fold.S b/src/stub/src/arm-linux.elf-fold.S
index 06774beb..8c5540c6 100644
--- a/src/stub/src/arm-linux.elf-fold.S
+++ b/src/stub/src/arm-linux.elf-fold.S
@@ -27,6 +27,7 @@
 @  <jreiser@users.sourceforge.net>
 @
 
+#define ARM_OLDABI 1
 #include "arch/arm/v4a/macros.S"
 
 sz_Elf32_Ehdr = 13*4
diff --git a/src/stub/src/armeb-linux.elf-entry.S b/src/stub/src/armeb-linux.elf-entry.S
index 8f9b7d4a..35fe9eee 100644
--- a/src/stub/src/armeb-linux.elf-entry.S
+++ b/src/stub/src/armeb-linux.elf-entry.S
@@ -29,7 +29,7 @@
 *  <jreiser@users.sourceforge.net>
 */
 
-
+#define ARM_OLDABI 1
 #include "arm-linux.elf-entry.S"
 
 
diff --git a/src/stub/src/armeb-linux.elf-fold.S b/src/stub/src/armeb-linux.elf-fold.S
index 35c861df..f1d515d2 100644
--- a/src/stub/src/armeb-linux.elf-fold.S
+++ b/src/stub/src/armeb-linux.elf-fold.S
@@ -27,6 +27,7 @@
 @  <jreiser@users.sourceforge.net>
 @
 
+#define ARM_OLDABI 1
 #define FILTER_ID 0x51  /* big-endian */
 #include "arm-linux.elf-fold.S"
 #undef FILTER_ID
diff --git a/src/stub/src/i386-darwin.macho-main.c b/src/stub/src/i386-darwin.macho-main.c
index 79f1d4af..3e2d00af 100644
--- a/src/stub/src/i386-darwin.macho-main.c
+++ b/src/stub/src/i386-darwin.macho-main.c
@@ -304,7 +304,6 @@ typedef union {
 #define PROT_WRITE     2
 #define PROT_EXEC      4
 
-typedef long long off_t;
 extern char *mmap(char *, size_t, unsigned, unsigned, int, off_t);
 extern ssize_t pread(int, void *, size_t, off_t);
 extern void bswap(void *, unsigned);
@@ -312,7 +311,7 @@ extern void bswap(void *, unsigned);
 static Mach_i386_thread_state const *
 do_xmap(
     Mach_header const *const mhdr,
-    unsigned const fat_offset,
+    off_t fat_offset,
     Extent *const xi,
     int const fdi,
     Mach_header **mhdrpp,
@@ -398,7 +397,7 @@ upx_main(
 )
 {
     Mach_i386_thread_state const *entry;
-    unsigned fat_offset = 0;
+    off_t fat_offset = 0;
     Extent xi, xo, xi0;
     xi.buf  = CONST_CAST(char *, 1+ (struct p_info const *)(1+ li));  // &b_info
     xi.size = sz_compressed - (sizeof(struct l_info) + sizeof(struct p_info));
diff --git a/src/stub/src/include/darwin.h b/src/stub/src/include/darwin.h
index 1d434162..6a564b53 100644
--- a/src/stub/src/include/darwin.h
+++ b/src/stub/src/include/darwin.h
@@ -69,6 +69,11 @@ typedef unsigned long long uint64_t;
 #endif
 typedef size_t uintptr_t;
 
+// XXX: restrict ourselves to 4GB for off_t.  Some versions of gcc
+// have bugs in handling 64-bit integers (such as passing as argument
+// the wrong registers) and it takes more code anyway.
+// Adjust in system call wrappers, particularly mmap() and pread().
+typedef unsigned off_t;
 
 // misc constants
 
diff --git a/src/stub/src/powerpc-darwin.macho-main.c b/src/stub/src/powerpc-darwin.macho-main.c
index ca049f2e..fe097c7a 100644
--- a/src/stub/src/powerpc-darwin.macho-main.c
+++ b/src/stub/src/powerpc-darwin.macho-main.c
@@ -313,15 +313,15 @@ typedef union {
 
 /* bug in crosstool/powerpc-750-linux-gnu/gcc-3.4.1-glibc-20040827:
    unsigned long long off_t goes into registers (9,10) instead of (8,9).
-   Adjust in mmap().
+   Adjust in mmap(), pread(), and include/darwin.h .
 */
-extern char *mmap(char *, size_t, unsigned, unsigned, int, /*off_t*/size_t);
-ssize_t pread(int, void *, size_t, /*off_t*/unsigned);  // FIXME?
+extern char *mmap(char *, size_t, unsigned, unsigned, int, off_t);
+ssize_t pread(int, void *, size_t, off_t);
 
 static Mach_ppc_thread_state const *
 do_xmap(
     Mach_header const *const mhdr,
-    unsigned const fat_offset,
+    off_t const fat_offset,
     Extent *const xi,
     int const fdi,
     Mach_header **mhdrpp,
@@ -399,7 +399,7 @@ upx_main(
 )
 {
     Mach_ppc_thread_state const *entry;
-    unsigned fat_offset = 0;
+    off_t fat_offset = 0;
     Extent xi, xo, xi0;
     xi.buf  = CONST_CAST(char *, 1+ (struct p_info const *)(1+ li));  // &b_info
     xi.size = sz_compressed - (sizeof(struct l_info) + sizeof(struct p_info));